[ODE] Parallelizing ODE?

[Kenneth Holmlund]

Hi,
Russel (and everyone else), have you given parallelization of ODE
a thought? We think it would be a cool idea to run ODE on the
64 dual P4 2Ghz Beowulf cluster that we are building right now and
make that feed a parallelized scenegraph on a 4 node graphics cluster.

This would require some profiling of the ODE core to check where
the bottlenecks are, and how well it actually can be parallelized, and 
if so,
what the requirements on bandwidth and latency would be.

In principle one could take collisions and dynamics into separate 
threads and
do some to magic to make that work asynchronously. However, we'd only gain
a factor from that but instead we would rather like to attack the N-body 
problem
with a large number och CPU:s in order to simulate really big systems.

I have been testing this idea with a company that is, err, related to 
ODE, but
the response we get is about the license fee for the closed source. That 
seems
to be a dead end, so we have decided to go for ODE instead.

Has anyone tried to compile ODE with the Intel compiler in order to check if
it can produce any SIMD code? Has anyone a clue about cache issues and ODE?
What's the typical size of the matrices for an N-body system (N= 10, 
100, 1000`, ...)?

/Kenneth

-- 
Kenneth Holmlund
Director VRlab
HPC2N, Umeå University, Sweden
holmlund at hpc2n.umu.se
T: +46-90-786 9655
C: +46-70-631 5520